1

Evaluation of applications over an intranet

2

FDDI100 MbpsLan 5

R1 R3

R2

R4

Lan 110 Mbps Eth

Lan 210 Mbps Eth

Lan 310 Mbps Eth

Lan 416 Mbps TR

File server 4

File server 2

Web server

100 Windows NT clients

File server 3120

Windows NT clients

File server 1

50 Unix Workstation

Intranet architecture

100 Windows NT clients

3

Workload analysis

Intranet applications (there different classes):

1) Corporate training: web based applicatin

2) Access to local file system

4

Infrastructure hypothesis

• All file server and web server have a single CPU and a single disk

• The FDDI and the router are very fast versus the Lan Ethernet, then can be modelled as a simple delay, therefore they are server without queue

• All the other components can be modelled as queues with service time independent by the load.

5

User hypotheses

• Finite number of users• An user generates a new request, after having received an answer to the previous request, after a time equal to its

thinking time• 85% of users is working with the local file system• 15% of users is working with the Web Server

6

Queueing network

Multiclass closed queue model, each class is charaterized by: client group, application, server

• client group: CLi: clients in Lan i (i: 1 to 4)• application:

– FS for local file server access,

– TR for Training

• server: – FSi: i-th NFS server (i: 1 to 4)

– WebS: Web Server

7

Tipi di classi e numero di utenti

(CL1, FS, FS1) 120 x 0.85 = 102

(CL2, FS, FS2) 50 X 0.85 = 43

(CL3, FS, FS3) 100 x 0.85 = 85

(CL4, FS, FS4) 100 x 0.85 = 85

(CL1, TR, WebS) 120 x 0.15 = 18

(CL2, TR, WebS) 50 x 0,15 = 7

(CL3, TR, WebS) 100 x 0,15 = 15

(CL4, TR, WebS) 100 x 0,15 = 15

8

Server types

Routers: low delay, given its low latency

FDDI ring: low delay, given high bandwidth

CPU

Disks servers with load independ service time

LANs

Service Demands: Di,r = Vi,r x Si,r

where Vi,r = Visit Ratio

Si,r= Service Time

9

FDDI R3R1

R4

R2

L2

D

D

D

D

C

C

CC

(CL1, Tr, Web)(CL2, Tr, Web)

(CL4, Tr, Web)(CL3, FS, Fs3)

L1 L3

(CL1,TR, Web)

(CL1, FS, Fs1)

(CL2,TR, Web) (CL2, FS, Fs2)

(CL3, TR, Web)

(CL1,TR, Web)(CL2,TR, Web)(CL3,TR, Web)(CL4,TR, Web)

(CL1, FS, Fs1)

QN model

Web S

10

Web server workload characterization

(for a training session)

• Avg request document size per HTTP request– 20 rqs for txt documents (2.000 bytes per doc)– 100 rqs for inline images (50.000 bytes each)

• (20 text pages x 5 inline/text pages)

– 15 rqs for other multi-media (mm) obj (2.000.000 bytes each)

11

Web server workload characterization

• % request for:– txt documents = 20/(20+100+15) = 15 %– inline images = 100/(20+100+15) = 74 %– other mm obj = 15/(20+100+15) = 11 %

12

Web server workload characterization

• Average document size0.15 x 2.000 + 0.74 x 50.000 + 0.11 x 2.000.000 =

= 257.300 bytes

Note it is an average, the overhead associated to the three document sizes is different

13

Web server workload characterization

• Document request arrival rate is function of the think time and of the number of users in the system

(CLi, TR, Web)#Usersi

think time

Max # Usersi

18 per Lan 17 per Lan 215 per Lan 315 per Lan 4

45 sec

14

Web server workload characterization

• Device service time– CPU: 1 msec processing time x HTTP request– Disk:

We need to consider• Seekrand= avg time to position at a random cylinder• DiskRevTIme = time for a complete disk revolution• TransferTime = BlockSize/ 106 x TransferRate

• ControllerTime = time spent at the controller for an I/O req.

Sd = ControllerTime +Pmiss x (SeekRand + DiskRevolutionTime/2+TransferTime)

15

Web server workload characterization

• Lan hp: no fragmentation i.e. max data area 1500 bytes; hp no data overhead for HTTP request

NDatagrams =

ServiceTimen =

MessageSize + TCPOvhd

minn MTUn - IPOvhd

Overheadn = TCPOvhd+Ndatagrams x (IPOvhd + FrameOvhdn)

8 x (MessageSize + Overheadn )

106 x Bandwidth

16

Web server workload characterization

• Lan hp: no fragmentation i.e. max data area 1500 bytes; hp no data overhead for HTTP request

• Ethernet

NDatagrams =

ServiceTimen =

257300 + 20

1500 - 20

Overheadn = 20 + Ndatagrams x (20 + 18)

8 x (257300 + 18 )

106 x Bandwidth

17

Web server workload characterization

• Lan hp: no fragmentation i.e. max data area 1500 bytes; hp no data overhead for HTTP request

• Token ring

NDatagrams =

ServiceTimen =

257300 + 20

1500 - 20

Overheadn = 20 + Ndatagrams x (20 + 28)

8 x (257300 + 28 )

106 x Bandwidth

18

Web server workload characterization

• Router• delay 134 sec x packet (approximated in total to 1 msec)

• FDDI• delay with

ServiceTimen =8 x (MessageSize + Overheadn )

106 x Bandwidth

19

Local file system workload characterization

• File dimension 8192 bytes• avg NFS request arrival rate is function of the think time and number od users in the system

(CLi, FS, FSi)#Usersi

think time

Max #Usersi

102 per Lan 143 per Lan 285 per Lan 385 per Lan 4

10 sec

20

Local file systemworkload characterization

• Device service time– CPU: 1 msec per file request– Disk:

We need to consider• Seekrand= avg time to position at a random cylinder• DiskRevTIme = time for a complete disk revolution• TransferTime = BlockSize/ 106 x TransferRate

• ControllerTime = time spent at the controller for an I/O req .• N blocks to read = 8192/2048 = 4

– Lan i with 8192 bytes

21

Throughput & response time

Class Throughput Response time (req/sec) (sec)

• CL1, FS, FS1 10,12 0,08• CL2, FS, FS2 4,23 0,06• CL3, FS, FS3 8,44 0,08• CL4, FS, FS4 8,44 0,07• CL1, TR, WebS 0,34 8,58• CL2, TR, WebS 0,14 8,55• CL3, TR, WebS 0,28 7,96• CL4, TR, WebS 0,28 8,35

22

Throughput & response time

Max#Usersi

response time + think timeThroughput =